trinart_characters_19.2m_stable_diffusion_v1

Introduction

The trinart_characters_19.2m_stable_diffusion_v1 is a model based on Stable Diffusion v1, trained using approximately 19.2 million anime/manga style images. It is designed to achieve a balance between artistic versatility and anatomical precision. This model was previously part of the AI Novelist/TrinArt service, and a newer version, Derrida, is currently under experimentation.

Architecture

The model utilizes 8 NVIDIA A100 40GB GPUs for processing. A custom autoencoder has been employed, with separate training for the autoencoder and latent diffusion model, enhancing the output quality.

Training

The model underwent training with a substantial dataset of 19.2 million augmented images, followed by fine-tuning with 50,000 additional images. The training process focused on ensuring artistic style versatility and maintaining high anatomical quality.

Guide: Running Locally

1. Setup Environment: Ensure your system is equipped with necessary libraries and tools for running Stable Diffusion models.
2. Download Checkpoints: Obtain the latest version of the model checkpoint and the custom autoencoder.
3. Override State Dict: Modify your script to override the state_dict for the first_stage_model as the official script does not support alternative VAEs.
4. Execution: Run the model locally, using GPU acceleration for optimal performance.

Cloud GPUs: Consider using cloud services like AWS, Google Cloud, or Azure, offering access to powerful GPUs such as NVIDIA A100.

License

This model is licensed under CreativeML OpenRAIL-M, which provides guidelines for usage and distribution.